The compute credits offered by Google to European researchers via RAISE facilitate access to artificial intelligence, but they create a dependency that is paid for at the time of exit. The two usual interpretations of this initiative leave aside the question of the portability of the work once access ends.
On October 15, 2026, the European Union launched the RAISE contribution initiative, the European virtual institute dedicated to artificial intelligence research. The day before, Google DeepMind and Google Cloud had announced their contribution: up to 200 European researchers will be able to access advanced models and three million dollars in compute credits.
Two interpretations oppose each other. For some, these credits offer a chance to teams unable to finance this type of research on their own. For others, they illustrate the paradox of a Europe seeking sovereignty from an American provider. Both, according to the analysis, forget the decisive question: will researchers be able to continue, reproduce, or transfer their work when access ends?
Access is not mastery
In many laboratories, access to computing conditions the possibility to experiment and publish. The credits therefore produce immediate and useful results. However, they guide technical choices: teams gradually adopt the provider’s interfaces, libraries, formats, and experimentation pipelines, and skills are built around these tools. Dependency does not arise from a contractual constraint but from the accumulation of technically rational decisions made in a subsidized environment.
Free credit reduces the entry cost, but it says nothing about the exit cost, which only appears when credits are exhausted, during a price change, or the disappearance of a model. Migration then affects environments, pipelines (automated sequences of processing steps), skills, and sometimes the reproducibility of results. According to the analysis, cloud credits offered to startups and companies follow the same logic.
A question of scientific continuity
In research, being able to exit an infrastructure means being able to reproduce one’s results. Can an experiment conducted on a model that has become inaccessible, or on proprietary components, be verified or resumed by another team after the program ends? If the answer depends entirely on the initial provider, the issue touches on the independence of public research.
The analysis distinguishes three levels of portability. Portability is said to be declared when the contract provides for data export and the provider claims to comply with open standards. It is mobilizable when the institution has the skills, budgets, and infrastructure necessary to migrate. It is verified when a complete experiment has been transferred or reproduced elsewhere, at a given date, with a measured and documented result. According to this text, an untested reversibility clause or an unusable “open” format do not guarantee continuity.
Conditioning aid on a dependency analysis
Depriving researchers of these resources would be counterproductive, the analysis estimates. It rather proposes that any substantial provision of cloud credits or proprietary models to public research be accompanied by an analysis of digital dependencies, inspired by the data protection impact assessment (DPIA), without adopting its formalism. This analysis would notably address the question of full recovery of data, parameters, logs, and results, the rights the provider retains over scientific productions, and the full cost of continuing work after the credits.
RAISE could thus require projects benefiting from private resources to submit a proportionate and updated portability dossier, while institutions would include the exit cost in their initial evaluation. The analysis acknowledges that this would burden projects but considers it necessary to prevent temporary aid from producing lasting dependency.
Measuring sovereignty
The European Commission’s Joint Research Centre published on the same day data and an analytical framework for digital sovereignty, structured into four dimensions: governance, infrastructures and software, products and markets, skills and people. The Centre refuses to derive a single score, sovereignty being, according to the analysis, a set of capabilities to be examined separately and then verified over time.
